Last updated: July 2026
Key Takeaways
- DeepSeek's V4-Flash API entered public beta on July 31, 2026 as build 0731: the same 284B-parameter architecture, re-post-trained, with DeepSeek-reported agent scores that pass even V4-Pro-Preview. The model name and calling method are unchanged.
- The upgrade is API-only. The MIT-licensed weights on Hugging Face are still the April preview, so the version you can download and own is now the older brain. Same name, two different models.
- Mainline llama.cpp merged DeepSeek V4 support in July, so the preview already runs on 128GB-class machines through standard tooling — and the day 0731 weights are published, the local path is already paved.
DeepSeek-V4-Flash 0731 is the new build behind the deepseek-v4-flash API as of July 31, 2026 — a re-post-train of the same 284-billion-parameter model, with dramatically higher agent scores by DeepSeek's own measurements. The catch for anyone who cares about ownership: the upgrade is API-only, and the downloadable weights are still April's preview.
That split is the story the launch coverage is skipping. The name you call did not change; the model answering it did. Meanwhile the copy you can put on your own disk — the one no vendor can retrain, retire, or reroute — just became the older of the two. Here is what shipped, what the benchmark asterisks say, and what you can actually run today.
What DeepSeek Shipped on July 31
The official V4-Flash API is now in public beta, and per DeepSeek's API changelog, the calling method is unchanged: set the model name to deepseek-v4-flash and you get the latest build. The company is explicit that this is a post-training upgrade only. Build 0731 keeps the exact architecture and size of the preview — 284 billion total parameters with 13 billion active per token, and the one-million-token context window — and the announcement states that the change applies only to the Flash API. The V4-Pro API and the models behind DeepSeek's app and website are untouched for now.
Two integration changes ride along. The build natively supports OpenAI's Responses API format, and it is, in DeepSeek's words, specifically adapted for Codex. The company's docs also list no-code backend swaps for Claude Code, GitHub Copilot, and OpenCode. More on what that means, and what it costs you, below.
The Benchmark Leap, With the Asterisks Attached
DeepSeek's numbers describe a small model that suddenly out-agents its bigger sibling. On Terminal Bench 2.1, the reported jump is 61.8 to 82.7, past V4-Pro-Preview's 72.1, and the pattern repeats across the company's agent suite.
| Benchmark | Flash-Preview | Flash-0731 | Pro-Preview |
|---|---|---|---|
| Terminal Bench 2.1 | 61.8 | 82.7 | 72.1 |
| DeepSWE | 7.3 | 54.4 | 12.8 |
| Cybergym | 38.7 | 76.7 | 52.7 |
| Toolathlon-Verified | 49.7 | 70.3 | 55.9 |
| NL2Repo | 39.4 | 54.2 | 38.5 |
| DSBench-Hard | 25.8 | 59.6 | 31.1 |
All scores DeepSeek-reported, July 31, 2026, measured with the company's own evaluation harness — described in its footnotes as upcoming, meaning unreleased — in minimal mode. DSBench-Hard is a DeepSeek-internal test set. No independent verification existed at publication.
Read the asterisks as carefully as the deltas. Every number above comes from DeepSeek testing DeepSeek, on a harness nobody outside the company can run yet, and two of the nine benchmarks it published are internal sets the company built itself. DeepSeek's own table also shows Anthropic's Opus 4.8 still ahead of 0731 on every listed benchmark, with GLM-5.2 landing between them on several — the claim here is proximity at Flash's cost, not victory. None of this means the leap is fake. A 47-point climb on DeepSWE is not a rounding artifact, and DeepSeek's preview-era claims tracked reasonably against third-party indexes like Artificial Analysis. It means the leap is currently unverified, and the harness release is the thing to watch if you want it to become checkable.
The Part You Cannot Download (Yet)
Here is the fact almost no launch story mentions: 0731 exists only behind the API. DeepSeek's announcement says today's upgrade applies only to the V4-Flash API, and the Hugging Face repository tells the same story: its model card still opens by presenting "a preview version" of the V4 series, lists no 0731 build, and carries the April evaluation tables. We checked the repository twice on launch day. It was touched within hours of the announcement, but nothing on it says the weights changed, and the card's own text says they have not.
So the name deepseek-v4-flash now refers to two different models depending on where you meet it. Through the API, it is the 0731 re-post-train with the new agent behavior. On your own disk, it is the April preview: MIT-licensed, permanent, and now the weaker of the two. That inversion is worth sitting with. For three months, open-weights users had the same Flash as API users. As of this morning, the owned copy is a generation behind.
To DeepSeek's credit, this is the transparent version of a model swap: announced, versioned, benchmarked, dated. Compare that with the silent variety we documented when Fable 5's quiet safeguard degraded answers with no notice at all. But transparent or silent, both prove the same rule: on any hosted model, which brain answers you is the operator's decision, made on the operator's schedule. The only version whose behavior you control is the one you can download — and right now, that version is the old one. DeepSeek's track record of following previews with open releases makes 0731 weights plausible. Nothing published promises them.
The Local Plot Twist: The Runtime Grew Up
There is a second change the launch coverage misses, and it cuts the other way: running the V4 family at home got dramatically easier while nobody was looking. When we published our V4-Flash hardware reality check in June, no stable release of llama.cpp, Ollama, or LM Studio could load the V4 architecture, and every consumer run depended on experimental forks. That is no longer true. Mainline llama.cpp added DeepSeek V4 support in pull request 24162, with a follow-up fix for quantized KV caches merged July 7. The support PR's own test tables show the 2-bit build at 80.76GB — landing within a rounding error of the 81GB figure our June piece cited — generating about 6 tokens per second on NVIDIA's 128GB-class GB10, per the maintainer's published runs. Modest speed, real mainline support.
On top of that, Unsloth now publishes GGUF builds it reports as lossless: the model's MXFP4 expert weights repacked bit-for-bit, with 100 percent top-token agreement against the official checkpoint, by Unsloth's own measurement. The practical meaning is that the April preview genuinely runs on the hardware class our June piece identified — roughly 96GB of VRAM or a 128GB unified-memory machine — through standard tooling now, not science projects. We are updating that reality check to reflect the merge.
And it sets up the interesting scenario: the runtime is already paved. If DeepSeek publishes 0731 weights, local users get the agent-grade Flash the same day, on hardware an enthusiast can actually own. That is the mirror image of the Kimi K3 situation, where the weights exist and the consumer hardware does not. What runs on your memory tier today is mapped in our guide to what AI models actually run at home.
How to Use 0731 Today
If the agent gains matter to you now, the API is the only door, and DeepSeek made it deliberately easy to walk through. The model name is unchanged, which means existing deepseek-v4-flash integrations are already running 0731 whether their owners noticed or not. DeepSeek's Codex integration guide covers OpenAI's coding agent specifically: Codex talks to models through the Responses API, which 0731 now speaks natively, and one configuration covers the Codex CLI, desktop app, and IDE extension. Per the same docs, only Flash supports Codex today, with V4-Pro support expected in early August 2026. The docs also list no-code backend swaps for Claude Code, GitHub Copilot, and OpenCode.
The standing caveat applies with extra force to agent workloads. An agent loop does not send a prompt; it sends your repository, your shell output, and every intermediate step to someone else's server, at whatever pace the loop decides. If jurisdiction or data handling matters to you, third-party hosts already carry the build — OpenRouter listed deepseek-v4-flash-0731 on launch day — and our launch coverage of the V4 family walks the sovereignty options, from non-DeepSeek hosts to routing only non-sensitive traffic through the cheap endpoint.
What to Watch Next
Three signals decide how this story ages, and none of them requires trusting anyone's benchmarks. First: 0731 weights appearing on the Hugging Face repository. That is the moment the ownership gap closes and the local story flips from running yesterday's model to running the agent-grade Flash on day one. Second: the release of the DeepSeek Harness the company's own footnotes call upcoming, which would let third parties reproduce the launch numbers instead of quoting them. Third: the first independent evaluations of 0731 through the API. DeepSeek's docs already point at a fourth thing coming — V4-Pro Codex support in early August — which suggests the agent push is a family-wide roadmap rather than a one-off. We will update this page as each lands, with dates.
Frequently Asked Questions
Is DeepSeek V4-Flash 0731 open source?
Not yet. The 0731 build is API-only as of July 31, 2026, per DeepSeek's own announcement. The downloadable DeepSeek-V4-Flash weights on Hugging Face remain the April preview checkpoint, released under the MIT License. DeepSeek's pattern of following previews with open releases makes 0731 weights plausible, but none have been published and none have been promised.
Can you run DeepSeek V4-Flash 0731 locally?
No. Only the April preview weights are downloadable, so the 0731 agent improvements exist exclusively behind DeepSeek's API until new weights are published. The preview itself does run locally: mainline llama.cpp now supports the V4 architecture, and the realistic entry point is roughly 96GB of VRAM or a 128GB unified-memory machine.
What actually changed in the 0731 build?
Nothing structural. It is the same 284-billion-parameter mixture-of-experts model with 13 billion active per token, re-post-trained for agent work. DeepSeek reports large agent-benchmark gains — Terminal Bench 2.1 rising from 61.8 to 82.7 — plus native Responses API support and Codex adaptation. All published scores were measured by DeepSeek on its own unreleased harness.
Is the API model different from the Hugging Face model now?
Yes, as of July 31, 2026. Calling deepseek-v4-flash through the API serves the 0731 re-post-train, while the Hugging Face repository still hosts the April preview. Same name, two different checkpoints. It is the clearest recent example of why an API model's identity belongs to the operator: the brain behind the name changed overnight, with no action from users.
Can I still run the preview version at home?
Yes, and more easily than at any point since release. Mainline llama.cpp merged V4 support in July 2026, and Unsloth publishes GGUF quantizations it reports as bit-for-bit lossless against the official checkpoint. Our hardware reality check covers the memory math; the short version is that a 96GB GPU rig or a 128GB unified-memory machine is the working floor.
Does 0731 cost more than the preview?
DeepSeek's announcement and changelog make no mention of a price change; the model name and calling method are unchanged, and existing integrations picked up the new build automatically. Check DeepSeek's pricing page for current first-party rates, and third-party hosts like OpenRouter for theirs. Both move faster than articles do, which is exactly why we do not print them.
